What a logistics veteran with 12+ years of experience across multiple 3PL and e-commerce fulfillment environments learned from running a 7-unit AMR fleet in a real warehouse — and why "just add more robots" is never the answer.
The Operator's Perspective
Most conversations about warehouse automation happen at 30,000 feet — strategy decks, ROI projections, vendor brochures. Rarely do you hear from the person who actually stands on the warehouse floor when robots stop moving.
We recently spoke with a logistics operations manager at a major Japanese retailer — someone with over 12 years of hands-on experience across multiple third-party logistics (3PL) providers before taking on his current role as the automation project manager for an e-commerce fulfillment center. He has personally evaluated, selected, deployed, and operated autonomous mobile robots (AMRs) across three distinct use cases, all within his current organization.
His perspective is rare: he is both the operator who has to fix problems when they happen and the decision-maker who controls the budget for automation investments.
Three Use Cases, One Fleet
The facility operates 7 AMRs across approximately 500 tsubo (~1,650 m²) of dedicated e-commerce warehouse space. The robots serve three distinct functions:
| Use Case | Description |
|---|---|
| E-commerce order picking | The primary use case. WMS sends picking instructions to the robot control system. The AMR navigates to storage, displays item and quantity on screen, waits for a human picker, then returns automatically. |
| Regional sorting | Packed orders are loaded onto AMRs and transported to staging areas organized by delivery destination for last-mile carrier handoff. |
| Replenishment (inbound putaway) | New inventory is inspected, loaded onto AMRs, and transported to storage locations — eliminating the most time-consuming walking in the putaway process. |
The System Architecture: WMS to RCS to AMR
The data flow reveals both the power and the limits of the current setup.
Orders arrive from multiple e-commerce platforms. They flow into a third-party WMS package, which generates outbound shipping instructions. These instructions are passed to the AMR vendor's Robot Control System (RCS) via a middleware layer that translates WMS data into robot-understandable commands.
A critical detail: this integration required three-party coordination — the WMS vendor, the AMR vendor, and the retailer's own team — to define requirements and build the data bridge. This is one reason the company strongly prefers a single AMR vendor: adding a second would mean another round of costly WMS integration work.
Congestion: The Problem That Nearly Broke the Deployment
When the AMR fleet was first deployed, congestion was an almost daily occurrence.
The root cause was straightforward but devastating: the warehouse used fixed-location storage — each product had one designated shelf position. When a popular item sold heavily during a promotional event, multiple robots would all converge on the same location simultaneously. With only one narrow aisle to access each shelf, robots would queue up, block each other, and eventually freeze.
How Recovery Works
When a robot deadlock occurs, recovery is manual. A floor worker presses the emergency stop button on the stuck robot, physically pushes it out of the congested area, then releases the emergency stop to resume autonomous operation.
The operation is simple enough that any worker can perform it — no specialized technician is needed. But it interrupts productive work, and during peak periods, the cumulative disruption is significant.
What Causes Stoppages?
We asked directly: what is the most common cause of AMR stoppages?
The Fix: From Fixed Locations to Free Locations
The team's response was a fundamental change in warehouse operations: they switched from fixed-location storage to free-location (chaotic) storage — the same approach famously used by Amazon.
Instead of assigning each product to one specific shelf, incoming inventory is placed in whatever empty slot is available. The same product may be scattered across dozens of locations throughout the warehouse. The WMS tracks it all digitally.
The effect on congestion was dramatic:
But the Underlying Problem Remains
Here is what makes this finding important for anyone thinking about fleet scaling: the congestion was solved operationally — by changing how inventory is stored — not algorithmically. The robots' navigation logic did not improve. They still take the shortest path to every destination. With only 7 units and dispersed inventory, the odds of collision are low enough to be manageable.
But what happens at 15 robots? 30? 100?
The manager acknowledged this directly:
This is perhaps the most important statement in the entire interview. The operator who actually runs the fleet identified the exact technical limitation — shortest-path routing causing route convergence — and articulated the solution concept (dynamic detour routing) independently. And when he brought this to his AMR vendor, the answer was: we can't do it.
The Missing Layer: No Unified Control
We asked about how AMRs coexist with other equipment — conveyors, forklifts, sorters — on the same warehouse floor.
The AMRs operate within their mapped zones and have collision avoidance for humans. But forklifts have no system-level control at all — they are managed purely through operational rules: "You can only operate in this area." There is no shared position data, no priority system, no automated coordination between equipment types.
When pressed on how "traffic control" works today:
"It's rule-based operational management. We can't apply system-level control — at least not with our current setup."
This is area separation by policy, not by technology. It works at small scale because the zones do not overlap much. But it is a fundamentally manual approach that does not scale.
Productivity: 4x Improvement, with a Ramp-Up
| Metric | Before AMR | After AMR |
|---|---|---|
| Picking rate | 30 items/person/hour | 120 items/person/hour |
| Improvement | — | 4x |
| Ramp-up period | — | 3-4 months |
This improvement did not happen overnight. The ramp-up to target productivity took 3-4 months after go-live, during which the team was tuning processes, adjusting layouts, and training workers.
When Peak Demand Exceeds Capacity
Normal daily volume averages 500-600 orders. During sales events, this spikes to 1,400-1,500 orders per day. On several occasions, the AMR fleet simply could not keep up — orders had to be deferred to the next day.
The financial impact is nuanced. Direct revenue loss was avoided by setting customer delivery expectations conservatively (1-week window). But overtime labor costs to process the backlog were real and measurable.
Battery and Charging Dynamics
| Parameter | Value |
|---|---|
| Charge rate | ~1% per minute (full charge in ~1.5 hours) |
| Run time | ~7 hours per full charge |
| Auto-charge threshold | 30% remaining |
During peak periods when all 7 robots have been running continuously, charging queue congestion occurs — multiple robots attempting to return to the charging station simultaneously, creating a secondary bottleneck.
How They Buy: Inside Japanese Enterprise Procurement
This interview produced perhaps the most detailed account we have gathered of how a Japanese enterprise actually selects and procures warehouse automation technology.
The Selection Process
The process follows a structured funnel over approximately 6 months:
- ~10 vendors initially identified and invited to present
- Each vendor scored against standardized criteria using a point-based comparison matrix
- Shortlisted vendors' reference sites visited to see robots in actual operation
- Field narrowed to ~3 vendors for competitive bidding
- Final decision by authorized budget holder
The Four Selection Criteria
| Priority | Criterion | Details |
|---|---|---|
| 1 | Deployment track record | More installations mean more customization available for reuse and greater confidence in reliability |
| 2 | Initial cost | Upfront investment remains a primary concern |
| 3 | Running/maintenance cost | Ongoing operational expenses including maintenance contracts |
| 4 | Lead time | Deteriorated significantly: was 2-3 months, now 6 months to 1 year |
Budget Authority Structure
The decision-making hierarchy is clearly defined by investment amount. Section-level managers handle approvals in the single-digit millions of yen. Department heads can authorize investments up to a significantly higher threshold. Beyond that, executive-level approval is required.
ROI Methodology
Investment cases are built on two components:
- Labor cost reduction — direct headcount or hours saved
- Revenue opportunity preservation — increased throughput means fewer missed shipment cutoffs, protecting sales that would otherwise be lost
The payback calculation combines both savings and preserved revenue, divided by total investment, to produce a payback period in years.
Willingness to Invest in Software Improvements
Subsidies: A Significant Factor
Government subsidies play a material role in automation investment decisions. The company recently deployed a sorting robot system where 50% of the total investment was covered by government subsidies.
Multiple subsidy programs exist — national-level grants, municipal assistance programs, and tax incentive schemes. Importantly, these programs can cover both hardware and software components. The application process is described as "not a particularly high hurdle."
Market Trends: What's Changed in the Last 2-3 Years
When asked about shifts in the AMR landscape, the manager offered a surprising observation: AMR use cases are actually narrowing, not expanding. The market appears to be trending back toward more traditional automated storage systems.
The current demand leans heavily toward automated storage and retrieval systems (AS/RS), devanners, and unmanned forklifts — equipment that addresses physically demanding tasks humans find difficult. Robots that specifically relieve ergonomic strain on human workers are seeing the strongest demand.
The Data-Driven Warehouse
Layout optimization happens monthly. The team analyzes the previous month's order data, cross-references it with historical patterns ("last year, small items sold heavily in this season"), and adjusts storage locations, aisle configurations, and zone boundaries accordingly.
This level of continuous operational tuning is described as standard practice for 3PL operations — weekly adjustments are not uncommon at larger-scale operators.
The Detour Problem: Why Vendors Say "That's Difficult"
The most technically significant moment in our conversation came toward the end, when we shared examples of robot congestion from other environments — autonomous vehicles clustering in parking lots, restaurant service robots jamming in narrow aisles, warehouse robots entering infinite loops at intersections.
The manager immediately connected these examples to his own experience:
"Robots try to take the shortest distance. So they all end up on the same route."
He then articulated what he believed the solution should look like:
"If there were a concept of detour routing, I think congestion wouldn't happen."
And the painful reality:
"But when you ask for that, the answer is always: 'that's difficult.'"
He then offered unsolicited strategic advice: rather than approaching AMR manufacturers directly (who have incentives to keep customers locked into single-vendor solutions), the most promising path would be partnering with vendor-agnostic system integrators and 3PL operators who manage multi-vendor environments and feel this pain most acutely.
"If you go to a manufacturer, it becomes about the manufacturer's agenda."
Key Data Points
| Metric | Value | Source |
|---|---|---|
| Fleet size | 7 AMRs (constant operation) | Direct statement (interview) |
| Warehouse floor area | ~500 tsubo (~1,650 m²) | Direct statement (interview) |
| Monthly order volume | ~30,000 orders | Direct statement (interview) |
| Daily order volume (normal) | 500-600 orders | Direct statement (interview) |
| Daily order volume (peak/sale) | 1,400-1,500 orders | Direct statement (interview) |
| Pre-AMR picking rate | 30 items/person/hour | Direct statement (interview) |
| Post-AMR picking rate | 120 items/person/hour (4x) | Direct statement (interview) |
| Ramp-up to target productivity | 3-4 months | Direct statement (interview) |
| Congestion frequency (initial) | Nearly daily | Direct statement (interview) |
| Congestion frequency (after fix) | Near zero | Direct statement (interview) |
| Top stoppage cause | Congestion (overwhelmingly) | Direct statement (interview) |
| Battery: charge rate | ~1%/min (full in ~1.5 hrs) | Direct statement (interview) |
| Battery: run time | ~7 hours per full charge | Direct statement (interview) |
| Auto-charge threshold | 30% remaining | Direct statement (interview) |
| Vendor selection process | 10 to 3 to 1 (over ~6 months) | Direct statement (interview) |
| AMR lead time (current) | 6 months to 1 year | Direct statement (interview) |
| AMR lead time (previous) | 2-3 months | Direct statement (interview) |
| Government subsidy (recent case) | 50% of investment | Direct statement (interview) |
| Layout optimization frequency | Monthly | Direct statement (interview) |
| Unified equipment control | None ("nothing manages all of it") | Direct statement (interview) |
This interview was conducted as part of Rovnou's ongoing research into robot fleet coordination challenges in Japanese logistics environments. The interviewee's identity and employer have been anonymized at their request. All data points are sourced directly from the recorded interview transcript.
Rovnou is developing multi-agent path finding (MAPF) software that coordinates heterogeneous robot fleets without replacing existing fleet management systems. To learn more, visit rovnou.com.